Upload lora-scripts/sd-scripts/library/utils.py with huggingface_hub
Browse files
lora-scripts/sd-scripts/library/utils.py
ADDED
@@ -0,0 +1,266 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import logging
|
2 |
+
import sys
|
3 |
+
import threading
|
4 |
+
import torch
|
5 |
+
from torchvision import transforms
|
6 |
+
from typing import *
|
7 |
+
from diffusers import EulerAncestralDiscreteScheduler
|
8 |
+
import diffusers.schedulers.scheduling_euler_ancestral_discrete
|
9 |
+
from diffusers.schedulers.scheduling_euler_ancestral_discrete import EulerAncestralDiscreteSchedulerOutput
|
10 |
+
|
11 |
+
|
12 |
+
def fire_in_thread(f, *args, **kwargs):
|
13 |
+
threading.Thread(target=f, args=args, kwargs=kwargs).start()
|
14 |
+
|
15 |
+
|
16 |
+
def add_logging_arguments(parser):
|
17 |
+
parser.add_argument(
|
18 |
+
"--console_log_level",
|
19 |
+
type=str,
|
20 |
+
default=None,
|
21 |
+
choices=["DEBUG", "INFO", "WARNING", "ERROR", "CRITICAL"],
|
22 |
+
help="Set the logging level, default is INFO / ログレベルを設定する。デフォルトはINFO",
|
23 |
+
)
|
24 |
+
parser.add_argument(
|
25 |
+
"--console_log_file",
|
26 |
+
type=str,
|
27 |
+
default=None,
|
28 |
+
help="Log to a file instead of stderr / 標準エラー出力ではなくファイルにログを出力する",
|
29 |
+
)
|
30 |
+
parser.add_argument("--console_log_simple", action="store_true", help="Simple log output / シンプルなログ出力")
|
31 |
+
|
32 |
+
|
33 |
+
def setup_logging(args=None, log_level=None, reset=False):
|
34 |
+
if logging.root.handlers:
|
35 |
+
if reset:
|
36 |
+
# remove all handlers
|
37 |
+
for handler in logging.root.handlers[:]:
|
38 |
+
logging.root.removeHandler(handler)
|
39 |
+
else:
|
40 |
+
return
|
41 |
+
|
42 |
+
# log_level can be set by the caller or by the args, the caller has priority. If not set, use INFO
|
43 |
+
if log_level is None and args is not None:
|
44 |
+
log_level = args.console_log_level
|
45 |
+
if log_level is None:
|
46 |
+
log_level = "INFO"
|
47 |
+
log_level = getattr(logging, log_level)
|
48 |
+
|
49 |
+
msg_init = None
|
50 |
+
if args is not None and args.console_log_file:
|
51 |
+
handler = logging.FileHandler(args.console_log_file, mode="w")
|
52 |
+
else:
|
53 |
+
handler = None
|
54 |
+
if not args or not args.console_log_simple:
|
55 |
+
try:
|
56 |
+
from rich.logging import RichHandler
|
57 |
+
from rich.console import Console
|
58 |
+
from rich.logging import RichHandler
|
59 |
+
|
60 |
+
handler = RichHandler(console=Console(stderr=True))
|
61 |
+
except ImportError:
|
62 |
+
# print("rich is not installed, using basic logging")
|
63 |
+
msg_init = "rich is not installed, using basic logging"
|
64 |
+
|
65 |
+
if handler is None:
|
66 |
+
handler = logging.StreamHandler(sys.stdout) # same as print
|
67 |
+
handler.propagate = False
|
68 |
+
|
69 |
+
formatter = logging.Formatter(
|
70 |
+
fmt="%(message)s",
|
71 |
+
datefmt="%Y-%m-%d %H:%M:%S",
|
72 |
+
)
|
73 |
+
handler.setFormatter(formatter)
|
74 |
+
logging.root.setLevel(log_level)
|
75 |
+
logging.root.addHandler(handler)
|
76 |
+
|
77 |
+
if msg_init is not None:
|
78 |
+
logger = logging.getLogger(__name__)
|
79 |
+
logger.info(msg_init)
|
80 |
+
|
81 |
+
|
82 |
+
|
83 |
+
# TODO make inf_utils.py
|
84 |
+
|
85 |
+
|
86 |
+
# region Gradual Latent hires fix
|
87 |
+
|
88 |
+
|
89 |
+
class GradualLatent:
|
90 |
+
def __init__(
|
91 |
+
self,
|
92 |
+
ratio,
|
93 |
+
start_timesteps,
|
94 |
+
every_n_steps,
|
95 |
+
ratio_step,
|
96 |
+
s_noise=1.0,
|
97 |
+
gaussian_blur_ksize=None,
|
98 |
+
gaussian_blur_sigma=0.5,
|
99 |
+
gaussian_blur_strength=0.5,
|
100 |
+
unsharp_target_x=True,
|
101 |
+
):
|
102 |
+
self.ratio = ratio
|
103 |
+
self.start_timesteps = start_timesteps
|
104 |
+
self.every_n_steps = every_n_steps
|
105 |
+
self.ratio_step = ratio_step
|
106 |
+
self.s_noise = s_noise
|
107 |
+
self.gaussian_blur_ksize = gaussian_blur_ksize
|
108 |
+
self.gaussian_blur_sigma = gaussian_blur_sigma
|
109 |
+
self.gaussian_blur_strength = gaussian_blur_strength
|
110 |
+
self.unsharp_target_x = unsharp_target_x
|
111 |
+
|
112 |
+
def __str__(self) -> str:
|
113 |
+
return (
|
114 |
+
f"GradualLatent(ratio={self.ratio}, start_timesteps={self.start_timesteps}, "
|
115 |
+
+ f"every_n_steps={self.every_n_steps}, ratio_step={self.ratio_step}, s_noise={self.s_noise}, "
|
116 |
+
+ f"gaussian_blur_ksize={self.gaussian_blur_ksize}, gaussian_blur_sigma={self.gaussian_blur_sigma}, gaussian_blur_strength={self.gaussian_blur_strength}, "
|
117 |
+
+ f"unsharp_target_x={self.unsharp_target_x})"
|
118 |
+
)
|
119 |
+
|
120 |
+
def apply_unshark_mask(self, x: torch.Tensor):
|
121 |
+
if self.gaussian_blur_ksize is None:
|
122 |
+
return x
|
123 |
+
blurred = transforms.functional.gaussian_blur(x, self.gaussian_blur_ksize, self.gaussian_blur_sigma)
|
124 |
+
# mask = torch.sigmoid((x - blurred) * self.gaussian_blur_strength)
|
125 |
+
mask = (x - blurred) * self.gaussian_blur_strength
|
126 |
+
sharpened = x + mask
|
127 |
+
return sharpened
|
128 |
+
|
129 |
+
def interpolate(self, x: torch.Tensor, resized_size, unsharp=True):
|
130 |
+
org_dtype = x.dtype
|
131 |
+
if org_dtype == torch.bfloat16:
|
132 |
+
x = x.float()
|
133 |
+
|
134 |
+
x = torch.nn.functional.interpolate(x, size=resized_size, mode="bicubic", align_corners=False).to(dtype=org_dtype)
|
135 |
+
|
136 |
+
# apply unsharp mask / アンシャープマスクを適用する
|
137 |
+
if unsharp and self.gaussian_blur_ksize:
|
138 |
+
x = self.apply_unshark_mask(x)
|
139 |
+
|
140 |
+
return x
|
141 |
+
|
142 |
+
|
143 |
+
class EulerAncestralDiscreteSchedulerGL(EulerAncestralDiscreteScheduler):
|
144 |
+
def __init__(self, *args, **kwargs):
|
145 |
+
super().__init__(*args, **kwargs)
|
146 |
+
self.resized_size = None
|
147 |
+
self.gradual_latent = None
|
148 |
+
|
149 |
+
def set_gradual_latent_params(self, size, gradual_latent: GradualLatent):
|
150 |
+
self.resized_size = size
|
151 |
+
self.gradual_latent = gradual_latent
|
152 |
+
|
153 |
+
def step(
|
154 |
+
self,
|
155 |
+
model_output: torch.FloatTensor,
|
156 |
+
timestep: Union[float, torch.FloatTensor],
|
157 |
+
sample: torch.FloatTensor,
|
158 |
+
generator: Optional[torch.Generator] = None,
|
159 |
+
return_dict: bool = True,
|
160 |
+
) -> Union[EulerAncestralDiscreteSchedulerOutput, Tuple]:
|
161 |
+
"""
|
162 |
+
Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
|
163 |
+
process from the learned model outputs (most often the predicted noise).
|
164 |
+
|
165 |
+
Args:
|
166 |
+
model_output (`torch.FloatTensor`):
|
167 |
+
The direct output from learned diffusion model.
|
168 |
+
timestep (`float`):
|
169 |
+
The current discrete timestep in the diffusion chain.
|
170 |
+
sample (`torch.FloatTensor`):
|
171 |
+
A current instance of a sample created by the diffusion process.
|
172 |
+
generator (`torch.Generator`, *optional*):
|
173 |
+
A random number generator.
|
174 |
+
return_dict (`bool`):
|
175 |
+
Whether or not to return a
|
176 |
+
[`~schedulers.scheduling_euler_ancestral_discrete.EulerAncestralDiscreteSchedulerOutput`] or tuple.
|
177 |
+
|
178 |
+
Returns:
|
179 |
+
[`~schedulers.scheduling_euler_ancestral_discrete.EulerAncestralDiscreteSchedulerOutput`] or `tuple`:
|
180 |
+
If return_dict is `True`,
|
181 |
+
[`~schedulers.scheduling_euler_ancestral_discrete.EulerAncestralDiscreteSchedulerOutput`] is returned,
|
182 |
+
otherwise a tuple is returned where the first element is the sample tensor.
|
183 |
+
|
184 |
+
"""
|
185 |
+
|
186 |
+
if isinstance(timestep, int) or isinstance(timestep, torch.IntTensor) or isinstance(timestep, torch.LongTensor):
|
187 |
+
raise ValueError(
|
188 |
+
(
|
189 |
+
"Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to"
|
190 |
+
" `EulerDiscreteScheduler.step()` is not supported. Make sure to pass"
|
191 |
+
" one of the `scheduler.timesteps` as a timestep."
|
192 |
+
),
|
193 |
+
)
|
194 |
+
|
195 |
+
if not self.is_scale_input_called:
|
196 |
+
# logger.warning(
|
197 |
+
print(
|
198 |
+
"The `scale_model_input` function should be called before `step` to ensure correct denoising. "
|
199 |
+
"See `StableDiffusionPipeline` for a usage example."
|
200 |
+
)
|
201 |
+
|
202 |
+
if self.step_index is None:
|
203 |
+
self._init_step_index(timestep)
|
204 |
+
|
205 |
+
sigma = self.sigmas[self.step_index]
|
206 |
+
|
207 |
+
# 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
|
208 |
+
if self.config.prediction_type == "epsilon":
|
209 |
+
pred_original_sample = sample - sigma * model_output
|
210 |
+
elif self.config.prediction_type == "v_prediction":
|
211 |
+
# * c_out + input * c_skip
|
212 |
+
pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (sample / (sigma**2 + 1))
|
213 |
+
elif self.config.prediction_type == "sample":
|
214 |
+
raise NotImplementedError("prediction_type not implemented yet: sample")
|
215 |
+
else:
|
216 |
+
raise ValueError(f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`")
|
217 |
+
|
218 |
+
sigma_from = self.sigmas[self.step_index]
|
219 |
+
sigma_to = self.sigmas[self.step_index + 1]
|
220 |
+
sigma_up = (sigma_to**2 * (sigma_from**2 - sigma_to**2) / sigma_from**2) ** 0.5
|
221 |
+
sigma_down = (sigma_to**2 - sigma_up**2) ** 0.5
|
222 |
+
|
223 |
+
# 2. Convert to an ODE derivative
|
224 |
+
derivative = (sample - pred_original_sample) / sigma
|
225 |
+
|
226 |
+
dt = sigma_down - sigma
|
227 |
+
|
228 |
+
device = model_output.device
|
229 |
+
if self.resized_size is None:
|
230 |
+
prev_sample = sample + derivative * dt
|
231 |
+
|
232 |
+
noise = diffusers.schedulers.scheduling_euler_ancestral_discrete.randn_tensor(
|
233 |
+
model_output.shape, dtype=model_output.dtype, device=device, generator=generator
|
234 |
+
)
|
235 |
+
s_noise = 1.0
|
236 |
+
else:
|
237 |
+
print("resized_size", self.resized_size, "model_output.shape", model_output.shape, "sample.shape", sample.shape)
|
238 |
+
s_noise = self.gradual_latent.s_noise
|
239 |
+
|
240 |
+
if self.gradual_latent.unsharp_target_x:
|
241 |
+
prev_sample = sample + derivative * dt
|
242 |
+
prev_sample = self.gradual_latent.interpolate(prev_sample, self.resized_size)
|
243 |
+
else:
|
244 |
+
sample = self.gradual_latent.interpolate(sample, self.resized_size)
|
245 |
+
derivative = self.gradual_latent.interpolate(derivative, self.resized_size, unsharp=False)
|
246 |
+
prev_sample = sample + derivative * dt
|
247 |
+
|
248 |
+
noise = diffusers.schedulers.scheduling_euler_ancestral_discrete.randn_tensor(
|
249 |
+
(model_output.shape[0], model_output.shape[1], self.resized_size[0], self.resized_size[1]),
|
250 |
+
dtype=model_output.dtype,
|
251 |
+
device=device,
|
252 |
+
generator=generator,
|
253 |
+
)
|
254 |
+
|
255 |
+
prev_sample = prev_sample + noise * sigma_up * s_noise
|
256 |
+
|
257 |
+
# upon completion increase step index by one
|
258 |
+
self._step_index += 1
|
259 |
+
|
260 |
+
if not return_dict:
|
261 |
+
return (prev_sample,)
|
262 |
+
|
263 |
+
return EulerAncestralDiscreteSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample)
|
264 |
+
|
265 |
+
|
266 |
+
# endregion
|